Re: [dpdk-dev] [PATCH 5/7] net/hns3: fix location of saving ethdev

[Ferruh Yigit <ferruh.yigit@intel.com>]

On 12/13/2020 8:03 AM, Lijun Ou wrote:
> From: "Min Hu (Connor)" <humin29@huawei.com>
> 
> In current version, procedure of saving eth_dev in
> hns3 PMD init will be called more than twice, one
> for primary, the other for secondary. That will cause
> segmentation fault in Multi-process as eth_dev will
> be changed in secondary process, which is different
> from one in primary process.
> This patch saved eth_dev in dev_private only in
> primary process, as dev_private is shared by primary
> process and secondary process.
> 
> Fixes: 8929efbc1c46 ("net/hns3: fix FEC state query")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Min Hu (Connor) <humin29@huawei.com>
> Signed-off-by: Lijun Ou <oulijun@huawei.com>
> ---
>   drivers/net/hns3/hns3_ethdev.c    | 3 +--
>   drivers/net/hns3/hns3_ethdev_vf.c | 1 +
>   2 files changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/net/hns3/hns3_ethdev.c b/drivers/net/hns3/hns3_ethdev.c
> index 7c34e38..f49b779 100644
> --- a/drivers/net/hns3/hns3_ethdev.c
> +++ b/drivers/net/hns3/hns3_ethdev.c
> @@ -6106,8 +6106,6 @@ hns3_dev_init(struct rte_eth_dev *eth_dev)
>   
>   	PMD_INIT_FUNC_TRACE();
>   
> -	hns->eth_dev = eth_dev;
> -
>   	eth_dev->process_private = (struct hns3_process_private *)
>   	    rte_zmalloc_socket("hns3_filter_list",
>   			       sizeof(struct hns3_process_private),
> @@ -6134,6 +6132,7 @@ hns3_dev_init(struct rte_eth_dev *eth_dev)
>   		return 0;
>   	}
>   
> +	hns->eth_dev = eth_dev;
>   	eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
>   

The definition of the problem is correct, but with this update the 
'hns->eth_dev' will be wrong for the secondary process.

The initial problem was access to 'rte_eth_devices' global variable, which is 
wrong. But current approach can cause problem for the secondaries, moving 
'eth_dev' to process private can work but before making things more complex,
I would like to check why it is needed.

The usage is as following:

hns3_query_dev_fec_info(struct hns3_hw *hw)
	struct rte_eth_dev *eth_dev = hns->eth_dev;
	ret = hns3_fec_get(eth_dev, &pf->fec_mode);

hns3_fec_get(struct rte_eth_dev *dev, uint32_t *fec_capa)
	struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);


Here since 'hns3_fec_get()' is dev_ops, it needs to get 'eth_dev' as parameter. 
Trying to use it internally forces to know 'eth_dev', but instead it can be 
possible to create an internal function that gets "struct hns3_hw" as parameter 
and it can be called internally without knowing 'eth_dev', and the .dev_ops can 
be wrapper to this,
I mean something like this, what do you think:

hns3_fec_get_internal(struct hns3_hw *hw, uint32_t *fec_capa);

hns3_query_dev_fec_info(struct hns3_hw *hw)
	ret = hns3_fec_get_internal(hw, &pf->fec_mode);

hns3_fec_get(struct rte_eth_dev *dev, uint32_t *fec_capa)
	struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
	hns3_fec_get_internal(hw, fec_capa)


There is other patch to switch to 'hns->eth_dev', I wonder can same approach be 
applied to those usages?